← 研踪 PaperTrailarXiv:2610.32113 · v2
REACT: Representational Early Analysis and Control of Chain-of-Thought Trajectories
首次提交 2026-10-23 11:28
版本更新 2026-11-19 22:49
雷达观测 2026-11-20 14:52
摘要
Chain-of-thought (CoT) generation offers an opportunity to intervene before a reasoning model completes its answer. How does the safety effect of a single activation edit vary across early reasoning states, and what factors govern that variation? We study this question with REACT, a conditional activation intervention that learns a safety direction from paired safe and unsafe trajectories and applies it once to an early CoT hidden state. The model weights remain frozen, and all subsequent tokens are generated without further edits. To study timing, we train directions separately at each of the first five CoT positions and evaluate ungated edits with full post-edit regeneration. Across three DeepSeek-R1 distilled models, site-specific edits concentrate safety gains in the opening CoT tokens. Norm-matched controls on DeepSeek-R1-Distill-Llama-8B (8B) disentangle two factors behind this curve: holding an effective direction and edit budget fixed still yields much larger early (-) than later (-) gains, whereas the near-zero site-specific gain reflects higher optimisation variance despite rising risk-probe AUROC. Forced prefixes shift the high-gain region with the opening context, and a four-arm intervention shows that ~98% of the net gain at propagates through the changed next token. On the same 8B model, REACT raises the rate of fully safe CoT-answer outputs on StrongREJECT-273 from 29.06% to 82.05%, while keeping benign refusal and usefulness near baseline. An anonymized implementation is available at https://anonymous.4open.science/r/REACT_official2-333B.
作者备注:33 pages, 5 figures
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-19 22:49 · 最新观测版本
REACT: Representational Early Analysis and Control of Chain-of-Thought Trajectories
Chain-of-thought (CoT) generation offers an opportunity to intervene before a reasoning model completes its answer. How does the safety effect of a single activation edit vary across early reasoning states, and what factors govern that variation? We study this question with REACT, a conditional activation intervention that learns a safety direction from paired safe and unsafe trajectories and applies it once to an early CoT hidden state. The model weights remain frozen, and all subsequent tokens are generated without further edits. To study timing, we train directions separately at each of the first five CoT positions and evaluate ungated edits with full post-edit regeneration. Across three DeepSeek-R1 distilled models, site-specific edits concentrate safety gains in the opening CoT tokens. Norm-matched controls on DeepSeek-R1-Distill-Llama-8B (8B) disentangle two factors behind this curve: holding an effective direction and edit budget fixed still yields much larger early (-) than later (-) gains, whereas the near-zero site-specific gain reflects higher optimisation variance despite rising risk-probe AUROC. Forced prefixes shift the high-gain region with the opening context, and a four-arm intervention shows that ~98% of the net gain at propagates through the changed next token. On the same 8B model, REACT raises the rate of fully safe CoT-answer outputs on StrongREJECT-273 from 29.06% to 82.05%, while keeping benign refusal and usefulness near baseline. An anonymized implementation is available at https://anonymous.4open.science/r/REACT_official2-333B.
查看此版本解读 →查看此版本原文 ↗